Symbolic Attack Chain Generation from Atomic Red Team Techniques: An Empirical Study of Predicate Representation Granularity

This paper studies whether symbolic predicate granularity materially affects automated cyberattack-chain planning. Its pipeline uses an LLM to translate Atomic Red Team techniques into PDDL representations and Fast Downward for deterministic planning. The authors compare AURORA’s nine-category Attack Action Linking Model with an empirically reduced five-category scheme. Across a corpus of 16 techniques, 81.3% of outcomes were identical, while plan validity and cost were reported as largely insensitive to granularity. The main apparent benefit of the richer representation is finer structural detail in plan justification, not greater attack-chain viability.